Links for 2023-11-08
“How fast is AI improving? In this interactive explainer, explore how capable AI language models (LMs) like ChatGPT are in the past and present, to better understand AI’s future.” https://theaidigest.org/progress-and-dangers
Ultra-Long Sequence Distributed Transformer: Presents an efficient distributed training method for training transformers with long sequences. 5.6x faster and 10.2x more memory-efficient implementation compared to SotA sequence parallelism. Notable: Oak Ridge people on Summit. So, the protein & genetic LLM models finally got the attention of the supercomputing people? https://arxiv.org/abs/2311.02382
Open-source code for training locomotion controllers: These robots run on ice, gravel, over large stones, stairs, under obstacles, on slippery surfaces, dance, recover from falls & more! https://github.com/Improbable-AI/walk-these-ways
Combining LLMs with formal methods and mathematical modelling to verify cyber-physical AI systems, ultimately aiming to enabling globally transformative AI with provable safety. https://aria.org.uk/wp-content/uploads/2023/10/ARIA-Mathematics-and-modelling-are-the-keys-we-need-to-safely-unlock-transformative-AI-v01.pdf
Personas as a Way to Model Truthfulness in Language Models https://arxiv.org/abs/2310.18168
VR-NeRF: High-Fidelity Virtualized Walkable Spaces — “We present an end-to-end system for the high-fidelity capture, model reconstruction, and real-time rendering of walkable spaces in virtual reality using neural radiance fields.” https://vr-nerf.github.io/
Try the new OpenAI TTS API for free in this demo. https://huggingface.co/spaces/ysharma/OpenAI_TTS_Streaming (It can also speak German but with an American accent.)
China wants to win the gene therapy race—and it’ll spend millions https://www.technologyreview.com/2023/11/01/1082740/china-gene-therapty-deafness-hearing/ [https://archive.ph/wip/lpcak]
In a surprising finding, light can make water evaporate without heat https://news.mit.edu/2023/surprising-finding-light-makes-water-evaporate-without-heat-1031
The IQs of children who had been adopted were much more highly correlated with the IQs of their biological parents than with their adoptive parents — this was true both when they were teenagers and when they were adults. These findings provide further evidence for the predominance of genetic influences on adult intelligence over any other systematic source of variation. https://www.sciencedirect.com/science/article/abs/pii/S0160289621000635
"The PLA is struggling to attract enough technologically skilled recruits to operate all its modern weaponry. Many of those with the requisite skills are choosing higher-paid jobs in industry or, worse, going abroad. Tsinghua confers about 8,000 degrees each year. On average, according to the university, just 16 students a year have joined the PLA since 2005. Last year that number was 12." https://www.economist.com/special-report/2023/11/06/china-is-struggling-to-recruit-enough-highly-skilled-troops [https://archive.ph/uQJoh]